Large language models (LLMs) often exhibit complementary strengths. Model routing harnesses these strengths by dynamically directing each query to the most suitable model, given a candidate model pool. However, routing performance relies on accurate model representations, and adding new models typically requires retraining, limiting scalability. To address these challenges, we propose a novel routing method using in-context vectors to represent model capabilities. The method proceeds in two stages. First, queries are embedded and projected into vectors, with a projector and LLM-based router trained to reconstruct the original queries, aligning vector representations with the router’s semantic space. Second, each candidate model is profiled on a query set, and the router learns---based on in-context vectors of query and model performance---to predict whether each model can correctly answer new queries. Extensive experiments demonstrate that our method achieves state-of-the-art routing performance in both in-distribution and out-of-distribution tasks. Moreover, our method allows for seamless integration of new models without retraining the router.

Zhen Yang, Ping Jian, Zhongbin Guo, Zuming Zhang, Chengzhi Li, Yonghong Deng, Xinyue Zhang, Wenpeng Lu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Logical reading comprehension is a challenging task that entails grasping the underlying semantics of text and applying reasoning to deduce the correct answer. Prior researches have primarily focused on enhancing logical reasoning capabilities through Chain-of-Thought (CoT) or data augmentation. However, previous work constructing chain-of-thought rationales concentrates solely on analyzing correct options, neglecting the incorrect alternatives. Addtionally, earlier efforts on data augmentation by altering contexts rely on rule-based methods, which result in generated contexts that lack diversity and coherence. To address these issues, we propose a Premise-Oriented Data Augmentation (PODA) framework. This framework can generate CoT rationales including analyses for both correct and incorrect options, while constructing diverse and high-quality counterfactual contexts from incorrect candidate options. We integrate summarizing premises and identifying premises for each option into rationales. Subsequently, we employ multi-step prompts with identified premises to construct counterfactual context. To facilitate the model's capabilities to better differentiate the reasoning process associated with each option, we introduce a novel thought-path contrastive learning method that compares reasoning path between the original and counterfactual samples. Experimental results on three representative LLMs demonstrate that our method can improve the baselines substantially across two challenging logical reasoning benchmarks (ReClor and LogiQA 2.0).

Zhen Yang, Ping Jian, Chengzhi Li. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Implicit Discourse Relation Recognition (IDRR), which infers discourse logical relations without explicit connectives, is one of the most challenging tasks in natural language processing (NLP). Recently, pre-trained language models (PLMs) have yielded impressive results across numerous NLP tasks, but their performance still remains unsatisfactory in IDRR. We argue that prior studies have not fully harnessed the potential of PLMs, thereby resulting in a mixture of logical semantics, which determine the logical relations between discourse arguments, and general semantics, which encapsulate the non-logical contextual aspects (detailed in Sec.1). Such a mixture would inevitably compromise the logic reasoning ability of PLMs. Therefore, we propose a novel method that trains the PLMs through two semantics enhancers to implicitly differentiate logical and general semantics, ultimately achieving logical semantics enhancement. Due to the characteristic of PLM in word representation learning, these two semantics enhancers will inherently confront with each other, facilitating an augmentation of logical semantics by disentangling them from general semantics. The experimental results on PDTB 2.0 dataset show that the confrontation approach exceeds our baseline by 3.81% F1 score, and the effectiveness of the semantics confrontation method is validated by comprehensive ablation experiments.

Diffusion models have become a powerful generative modeling paradigm, achieving great success in continuous data patterns. However, the discrete nature of text data results in compatibility issues between continuous diffusion models (CDMs) and pre-trained language models (PLMs). That is, the performance of diffusion models even degrades when combined with PLMs. To alleviate this issue, we propose to utilize a pre-trained decoder to convert the denoised embedding vectors into natural language instead of using the widely used rounding operation. In this way, CDMs can be more effectively combined with PLMs. Additionally, considering that existing noise schedules in text diffusion models do not take into account the linguistic differences among tokens, which violates the easy-first policy for text generation, we propose a linguistic easy-first schedule that incorporates the measure of word importance, conforming to easy-first-generation linguistic features and bringing about improved generation quality. Experiment results on the E2E dataset and five controllable tasks show that our approach can combine the merits of CDMs and PLMs, significantly outperforming other diffusion-based models.

Most Seq2Seq neural model-based medical question summarization (MQS) systems have a severe mismatch between training and inference, i.e., exposure bias. However, this problem remains unexplored in the MQS task. To bridge this research gap and alleviate the problem of exposure bias, we propose a novel re-ranking training framework for MQS called Multi-view Contrastive Learning (MvCL). MvCL simultaneously considers the similarity scores between medical questions and candidate summaries as well as the average similarity scores between candidate summaries and other candidates within the same group, and utilizes contrastive learning to optimize the model’s ranking ability. Additionally, we propose a new multilevel inference approach to adapt to this training strategy. The approach first filters out candidate summaries that are dissimilar to the original medical question, and then selects the summary with the highest average similarity to other candidate summaries from the remaining candidates as the final output. We conducted extensive experiments, and the results demonstrate that our proposed MvCL framework achieves state-of-the-art results on the majority of evaluation metrics across four datasets.1

Logical reasoning tasks have recently become a research hotspot in machine reading comprehension communities. This task requires models to answer the question by extracting and utilizing the implicit logical information hidden in the text. Logical information includes logical content information and logical pattern information. When the context and options are relevant in content aspect, such as Necessary Assumption, logical content information can help model to perform better. When it comes to content unrelated situation, such as Logic Principle questions, models need to further distill logical pattern information. Previous works has proposed some strategies, focusing on modeling the logical content information, but there are still some limitations such as long-distance dependency, heavy reliance on external data and internal data augmentation. In addition, the logical pattern information has not received much attention in the previous works, which will cause negative impact on the generalization in practical scenarios. In this paper, we try to improve both effectiveness and generalization of the model. We proposed a novel logical Transformer Capsule Network (LTCN). In this model, to better capture the logical content information, we combine logic graph with Transformer by using biaffine mechanism. And we fill the gap of ignoring logical pattern information by introducing a Capsule Network. Experimental results shows that our model outperforms on both ReClor and LogiQA datasets. Specially, our model has a significant performance improvements on handling more challenging logical reasoning questions.

Image paragraph captioning aims to describe given images by generating natural paragraphs. Unfortunately, the paragraphs generated by existing methods typically suffer from poor coherence since the visual information is inevitably lost after the pooling operation, which maps numerous visual features to only one global vector. On the other hand, the pooled vectors make it harder for the language models to interact with details in images, leading to generic or even wrong descriptions of visual details. In this paper, we propose a simple yet effective module called Visual Information Enhancement Module (VIEM) to prevent the visual information loss in visual features pooling. Meanwhile, to model the inter-sentence dependency, a fusion gate mechanism, which makes the most of the nonpooled features by fusing visual vectors with textual information, is introduced into the language model to furthermore improve the paragraph coherence. In experiments, the visual information loss is quantitatively measured through a mutual information based method. Surprisingly, the results indicates that such loss in VIEM is only approximately 50% of that in pooling, effectively demonstrating the efficacy of VIEM. Moreover, extensive experiments on Stanford image-paragraph dataset show that the proposed method achieves promising performance compared with existing methods1.

Satellite images time series (SITS) change detection (CD) provides an efficient way to simultaneously access the temporal and spatial information about the observed region on the earth. However, the outputs of traditional SITS CD methods which are either binary maps or semantic change maps are often difficult to interpret by end users. And conventional remote sensing image change caption methods can only describe bi-temporal images. We propose SITS change caption, which not only identifies the changed regions in SITS but also summarize changes across SITS in natural language. Unfortunately, the scarcity of available SITS training datasets poses a major challenge for SITS change caption. To address these issues, this letter presents an innovative approach that leverages only bi-temporal remote sensing image change caption training data instead of SITS training data for SITS change captioning. Experimental results on real SITS dataset demonstrate the effectiveness of our proposed method, achieving better performance on all indicators. The observed improvements exceeded 20%. The source code can be downloaded from https://github.com/Crueyl123/SITSCC.

Deep learning based hyperspectral image (HSI) change detection (CD) has been a research hotspot. However, the high dimensionality and the limited training samples make HSI-CD difficult to implement. Besides, the data uncertainty inherited from HSIs is often neglected. To deal with these issues, this paper presents an uncertainty aware graph self-supervised learning (UA-GSSL) approach for unsupervised HSI-CD, which allows encoding not only the spectral and topological structure attributes but also the data uncertainty into learned feature representation for downstream CD task. Specifically, spectral and spatial correlations in HSIs are firstly characterized via graph model. Then, based on the constructed spectral-spatial graph models, novel node-level and edge-level data augmentations are devised to enrich the contrastive sample pairs. Thirdly, a graph based dual-branch self-supervised learning (SSL) contrastive network is designed to maximize the mutual information between a pair of low-dimensional feature embeddings. Fourthly, a simple but effective uncertainty aware loss function is dedicatedly formed to encourage the reliable features to play the more dominant role in feature representation. Finally, change map is produced using the learned features. Experimental results obtained on four real HSI datasets sufficiently demonstrate that the UA-GSSL achieves remarkable results compared to twelve state-of-the-art (SOTA) methods. The source code of this article can be downloaded from https://github.com/vickyiiiii /UA-GSSL.

Implicit Discourse Relation Recognition (IDRR), which infers discourse relations without the help of explicit connectives, is still a crucial and challenging task for discourse parsing.Recent works tend to exploit the hierarchical structure information from the annotated senses, which demonstrate enhanced discourse relation representations can be obtained by integrating sense hierarchy.Nevertheless, the performance and robustness for IDRR are significantly constrained by the availability of annotated data.Fortunately, there is a wealth of unannotated utterances with explicit connectives, that can be utilized to acquire enriched discourse relation features.In light of such motivation, we propose a Promptbased Logical Semantics Enhancement (PLSE) method for IDRR.Essentially, our method seamlessly injects knowledge relevant to discourse relation into pre-trained language models through prompt-based connective prediction.Furthermore, considering the prompt-based connective prediction exhibits local dependencies due to the deficiency of masked language model (MLM) in capturing global semantics, we design a novel self-supervised learning objective based on mutual information maximization to derive enhanced representations of logical semantics for IDRR.Experimental results on PDTB 2.0 and CoNLL16 datasets demonstrate that our method achieves outstanding and consistent performance against the current state-of-the-art models. 1

Implicit discourse relation recognition (IDRR), which infers discourse logical relations without the help of explicit connectives, is the bottleneck of discourse parsing. It is also an effective mean to test how well the natural language understanding models grasp the logical semantics of the text. Unfortunately, as an important part of text logical semantics, numerical logic has not been paid any attention to in the community. In this work, we attach importance to numerical semantics and design a numerical logic reasoning module specifically for the numeric tokens in discourse arguments to enhance the discourse logic inferring. Graph neural network is utilized here to calculate the interactions of these numerical elements by self-attention and inter-attention according to their numerical type and their location in the discourse arguments. Experimental results show that our model outperforms the baseline 1.36 % F1 score on the PDTB2.0 dataset.

It is hard to evaluate translations objectively and accurately, which limits the applications of machine translation. In this article, we assume that the above phenomenon is caused by noise interference during translation evaluation, and we handle the problem through a perspective of causal inference. We assume that the observable translation score is affected by the unobservable true translation quality and some noise simultaneously. If there is a variable that is related to the noise and independent to the true translation quality, the related noise can be eliminated by removing the effect of that variable from the observed score. Based on the above causality hypothesis, this article studies the length bias problem of beam search for neural machine translation (NMT) and the input related noise problem of translation quality estimation (QE). For the NMT length bias problem, we conduct the experiments on four typical NMT tasks (Uyghur–Chinese, Chinese–English, English–German, and English–French) with different scales of datasets. Comparing with previous approaches, the proposed causal motivated method is model-agnostic and does not require supervised training. For QE tasks, we conduct the experiments on the WMT’20 submissions. Experimental results show that the denoised QE results gain better Pearson’s correlation scores with human assessed scores compared to the original submissions. Further analyses on the NMT and QE tasks also demonstrate the rationality of the empirical assumptions made on our methods.

Currently, most of the supervised change detection approaches require a training data set that contains samples from both the changed and the unchanged data. However, under certain condition, such as natural disaster and military attack, the changed data samples are very few or even not available but the unchanged data are abundant. In this letter, we develop a generative adversarial networks (GANs)-based one-class classification (OCC) technique for time series remote-sensing image change detection. The proposed method is only trained with the unchanged data instead of both the changed and unchanged data. To achieve this purpose, first, spatial-spectral features are extracted from the time series remote-sensing images. Second, a GAN model is trained to detect the changes only with the extracted features of unchanged data. Remarkably, to offset the outlier errors caused by the incomplete supervision information provided by unchanged data alone, changed data, instead of unchanged data, are generated to improve power of discriminator. Finally, testing data are classified by the trained discriminator of GAN to produce a binary change map. Experimental results obtained on two optical time series remote-sensing data sets confirmed the effectiveness of our proposed method.

Statistical machine translation (SMT) models rely on word-, phrase-, and syntax-level alignments. But neural machine translation (NMT) models rarely explicitly learn the phrase- and syntax-level alignments. In this article, we propose to improve NMT by explicitly learning the bilingual syntactic constituent alignments. Specifically, we first utilize syntactic parsers to induce syntactic structures of sentences, and then we propose two ways to utilize the syntactic constituents in a perceptual (not adversarial) generator-discriminator training framework. One way is to use them to measure the alignment score of sentence-level training examples, and the other is to directly score the alignments of constituent-level examples generated with an algorithm based on word-level alignments from SMT. In our generator-discriminator framework, the discriminator is pre-trained to learn constituent alignments and distinguish the ground-truth translation from the fake ones, while the generative translation model is fine-tuned to receive the alignment knowledge and to generate translations that best approximate the true ones. Experiments and analysis show that the learned constituent alignments can help improve the translation results.

Chinese semantic matching is a fundamental task in natural language processing, which is critical and yet challenging for a series of downstream tasks. Although recent work on text representation learning has shown its potential in improving the performance on semantic matching, relatively limited work has been done on exploring the relevant interactive information between two granularity of Chinese text, i.e., character and word. Existing methods usually focus on capturing the interactive features from single granularity, which lead to inefficient text representation. Also, they typically fail to consider the fusion of features from different granularity. As a result, they only achieve limited performance improvement. This paper proposes a novel Chinese semantic matching model based on multi-granularity alignment and feature fusion (MAFFo). To be specific, we first encode the texts from different granularity, which are further handled with soft-alignment attention mechanism to extract relevant interactive information between texts on different granularity. In addition, we devise a feature fusion structure to merge the features from different granularity to generate an ideal representation for the pair of input text sequences, followed by a sigmoid function to judge the semantic matching degree. Extensive experiments on the publicly available dataset BQ demonstrate that our model can effectively improve the performance of semantic matching task and achieve comparable performance with BERT-based methods.

Yingxue Zhang, Fandong Meng, Peng Li, Ping Jian, Jie Zhou. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

The development of cognitive robotics brings an attractive scenario where humans and robots cooperate to accomplish specific tasks. To facilitate this scenario, cognitive robots are expected to have the ability to interact with humans with natural language, which depends on natural language understanding ( NLU ) technologies. As one core task in NLU, sentence semantic matching ( SSM ) has widely existed in various interaction scenarios. Recently, deep learning–based methods for SSM have become predominant due to their outstanding performance. However, each sentence consists of a sequence of words, and it is usually viewed as one-dimensional ( 1D ) text, leading to the existing available neural models being restricted into 1D sequential networks. A few researches attempt to explore the potential of 2D or 3D neural models in text representation. However, it is hard for their works to capture the complex features in texts, and thus the achieved performance improvement is quite limited. To tackle this challenge, we devise a novel 3D CNN-based SSM ( 3DSSM ) method for human–robot language interaction. Specifically, first, a specific architecture called feature cube network is designed to transform a 1D sentence into a multi-dimensional representation named as semantic feature cube. Then, a 3D CNN module is employed to learn a semantic representation for the semantic feature cube by capturing both the local features embedded in word representations and the sequential information among successive words in a sentence. Given a pair of sentences, their representations are concatenated together to feed into another 3D CNN to capture the interactive features between them to generate the final matching representation. Finally, the semantic matching degree is judged with the sigmoid function by taking the learned matching representation as the input. Extensive experiments on two real-world datasets demonstrate that 3DSSM is able to achieve comparable or even better performance over the state-of-the-art competing methods.

Sentence intention matching is vital for natural language understanding.Especially for Chinese sentence intention matching task, due to the ambiguity of Chinese words, semantic missing or semantic confusion are more likely to occur in the encoding process.Although the existing methods have enriched text representation through pre-trained word embedding to solve this problem, due to the particularity of Chinese text, different granularities of pre-trained word embedding will affect the semantic description of a piece of text.In this paper, we propose an effective approach that combines charactergranularity and word-granularity features to perform sentence intention matching, and we utilize soft alignment attention to enhance the local information of sentences on the corresponding levels.The proposed method can capture sentence feature information from multiple perspectives and correlation information between different levels of sentences.By evaluating on BQ and LCQMC datasets, our model has achieved remarkable results, and demonstrates better or comparable performance with BERT-based models.

Neural Machine Translation (NMT) optimized by Maximum Likelihood Estimation (MLE) lacks the guarantee of translation adequacy.To alleviate this problem, we propose an NMT approach that heightens the adequacy in machine translation by transferring the semantic knowledge learned from bilingual sentence alignment.Specifically, we first design a discriminator that learns to estimate sentence aligning score over translation candidates, and then the learned semantic knowledge is transfered to the NMT model under an adversarial learning framework.We also propose a gated self-attention based encoder for sentence embedding.Furthermore, an N -pair training loss is introduced in our framework to aid the discriminator in better capturing lexical evidence in translation candidates.Experimental results show that our proposed method outperforms baseline NMT models on Chinese-to-English and English-to-German translation tasks.Further analysis also indicates the detailed semantic knowledge transfered from the discriminator to the NMT model.

Ruiying Geng, Binhua Li, Yongbin Li, Xiaodan Zhu, Ping Jian, Jian Sun. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

This article presents a mutual learning method using hierarchical deep semantics for the classification of implicit discourse relations in English. With the absence of explicit discourse markers, traditional discourse techniques mainly concentrate on discrete linguistic features in this task, which always leads to a data sparseness problem. To relieve this problem, we propose a mutual learning neural model that makes use of multilevel semantic information together, including the distribution of implicit discourse relations, the semantics of arguments, and the co-occurrence of phrases and words. During the training process, the predicting targets of the model, which are the probability of the discourse relation type and the distributed representation of semantic components, are learned jointly and optimized mutually. The experimental results show that this method outperforms the previous works, especially in multiclass identification attributed to the hierarchical semantic representations and the mutual learning strategy.

This letter presents a hierarchical semantic multi-instance multilabel learning (MIML) framework for high-resolution (HR) remote sensing image annotation via Gaussian process (GP). The proposed framework can not only represent the ambiguities between image contents and semantic labels but also model the hierarchical semantic relationships contained in HR remote sensing images. Moreover, it is flexible to incorporate prior knowledge in HR images into the GP framework which gives a quantitative interpretation of the MIML prediction problem in turn. Experiments carried out on a real HR remote sensing image data set validate that the proposed approach compares favorably to the state-of-the-art MIML methods.